alt Hacker NewsShow HN: What should the GUI for AI agents look like?
Hi HN! We’re Akilan and Miguel, the creators of MarbleOS.
The inspiration for Marble comes from the GUI work at Xerox PARC, the 1984 Macintosh, and later NeXTSTEP, which became the foundation for Mac OS X. Before GUIs, interacting with a computer was limited to strange terminal commands:
C:\> DIR
C:\> COPY FILE.TXT A:
You had to remember the command, syntax, paths, and parameters.
The GUI made those capabilities visible. Instead of remembering commands, you could point at files, drag them, click buttons, and select actions from menus. It didn't necessarily make entirely new things possible; it just made existing capabilities much easier to understand and use. We feel like AI is still somewhere around this command-line stage.
Even though the strict syntax has been replaced with natural language, the interaction can still be quite stiff and depend on heavily recall. Tools like Claude Cowork still look surprisingly terminal-like: /skill-name [param1] [param2]. The parameters are written in natural language, but the user still needs to know the capabilities that exist, and how to invoke it. Command-line flags and arguments have just been replaced with tools, skills, and context.
ChatGPT works very well for asking questions, but we aren't convinced that it's the final interface for delegating work across multiple agents. A blank text box and a list of chat threads feels limiting. An agent's capabilities are mostly invisible. With Marble, we're exploring an interface that treats AI more like a workspace than another chat app. Each delegated task becomes a card. Multiple jobs can sit next to each other and run at once. Files, tools, and finished artifacts are visible at once. Before a task runs, Marble also shows which tools it expects to use. The basic idea is that the user should not have to hold the entire structure of the task in their head. And the result should be something directly usable, like a spreadsheet, PowerPoint, or other file, rather than something buried in a transcript.
Marble is our attempt at exploring interaction models beyond the chat box. The product is mainly for people who already use ChatGPT or Claude, but haven’t really adopted agent workflows yet. We’ve found that when the tools and possible actions are made more visible, people start delegating work they would not have thought to do through normal chat.
The site includes a downloadable beta if you want to try it. Does Marble feel like a genuinely novel interface?
Comments
When I initially saw this I was excited because what I’m trying to solve for is visibility of both semi and fully autonomous and agents at runtime. Marble is an improvement over chat and Claude but it’s incremental IMO.
We’ve toyed with a meta agent style interface which you chat to do everything. Add tools, connect datasets, build agents, deploy and schedule them, build skills, etc.
But at some point agents have to have an interface for human approvals, validation, training, and tuning.
I guess it’s like a spreadsheet. Ultimately agents are part of custom apps which are each going to be as unique like snowflakes. So it’s difficult to have one standard abstracted runtime interface without feeling obtuse to a business user. I think what I’m describing and wanting are new app layer building blocks vs one interface to rule them all.
All of this said, I’m all for a better personal llm harness interface (which is what this is). So keep up the great work guys!
I completely agree with you that the chat interface is very undercooked and there's a big untapped space for agent-first interfaces, perhaps even replacing traditional desktop metaphors. You're spot on with your premise.
With that said, from the material you show in your homepage, it still seems like you're doing something very similar to OpenAI and Anthropic's first party apps? I think it would help your pitch if you could show some task or workflow where your interface truly makes a difference; loading files and asking questions about them is a staple of most harnesses.
Best interface in my opinion is a git tracked folder, files as state, agents coming in and doing work. It's not just a chat interface because in chat mode everything flows like water under the bridge. It is files as agents, where you get the agent to write a task in a file, and then it updates the file like a blackboard as it performs work. Same task file can be seen by worker agent, multiple judge agents (for plan and implementation), documentation agents and reflexion agents (improve workflow). I would also track every user message in a chat_log.md for reflexion agents (what is wrong with the workflow) and intent review (is the agent still following user intent?).
I think it should look like this: https://www.dreamcoder.ai/
A smart, tiling window manager with workspaces and a great tool harness. An everything app, if you may; bring the terminals, code reviews, etc. into the tool itself, and make it high fidelity: no blocky terminals for the app itself.
Lost me at having to click to use particular tools
These Todo items on the side in the video seem to be useful.
I'm not so sure about that tool toolbar. It might as well just be a tool selector dropdown in a typical chat window.
A quick FYI: at marbles.com/learn, starting from „Turn Messy Notes into Today’s Task List“ the demo videos don’t work for me. I tried Safari and Chrome.
This is definitely an iterative improvement on what we have today. However, I think from a fundamental approach, you've just highlighted the pain of working with AI that I hadn't quite been able to verbalise until seeing this.
Go and have a look at the history of Microsoft OLE and OpenDoc and you'll see some of these ideas have been around for a while. When these ideas were first being shared in the industry (and yes, I am that old), here's the story that was being told:
In the future, you wouldn't buy "a word processor" or "an image editor", you'd buy components that could do specific tasks, and you'd bring them to your document/work as and when you needed. I'd have DTP components for layout, word processing components for spell checking, image components for resizing and colour balancing, and I'd be heads down in my fancy newsletter or report, and I'd be bringing functionality to the work I needed to do, without having to context switch and import/export things by hand.
The demo you've shown on the homepage is almost the opposite of this: I'm not bringing things into my work (fanning in), I'm starting agents to go and do lots of different bits of work (fanning out), that eventually I'm going to need to think about how to structure and fan in again.
I think coding interfaces kind of force us into a fan-in: here's my code base, now I want to improve test coverage, now I want to refactor, now I want to add this SDK, now I want to change this logic to use the SDK based on configuration params, now I want to improve my integration tests, and so on. I'm flowing through my work in a single asset (the code base), and able to bring agents and other tools to it.
Now, I'd like that, but for everything else. The UI I'd like allows me to build around the assets I am building.
Planning a trip? OK, I need to bring various agents along to help me plan an itinerary (in my calendar), based on location, budget, and preferences. It's not good enough to list the top 10 must-see places in Venice, I need them to fit into my calendar including travel time, and my budget - I want agents to help me plan that out. Ideally I want agents to help me and my partner both contribute to that based on our wants.
Need to do a report for the big boss? OK, I need to start with an empty report doc, and bring various sources of data, and take it through transformations to build the narrative/deck I need to write. Here's some excel data, here's an agent that can do some competitive analysis, here's another that is pulling quotes from real customers on social media, and so on...
Want to learn a new skill? We're going to need to have a framework for assessing where I am in terms of that skill (the "single asset" we're working on is "what I know"), and then agents to figure out the best way for me to fill gaps and reinforce what I already know.
Your demo is nicer than a chat box, but I think you need to think about the job that needs to be done, and how to bring agents to that job (perhaps even suggest those jobs - that could be another agent), not build a UI to just allow me to fire off dozens of agents that then have output sat in multiple windows across my screen.
It does feel novel.
However, I believe the GUI for AI agents that will win is a plain text editor with a file tree, a terminal pane for coding (or preview pane for other kinds of tasks) and a chat sidebar.
Even if the code editor is not used to type, it's there for psychological safety. It lets you inspect and navigate what is being created. Has tremendous value.
Plain text has survived countless software revolutions, so it's likely to become the dominant format for artifacts produced by AI. Non-plain text things are likely to adapt instead of the other way around.
"Strange terminal commands" have survived, while even poweruser toolbar style have started to fade (I love them, but they're fading away). It clearly has some value (hint: speed over familiarity).
I think the poweruser toolbar style was indeed the only thing in the past couple of decades that really tried to match the terminal for speed of interaction. In my opinion, it succeeded, but it has its limits too (tends to clump with stuff over time; has a limit on the amount of abstractions it can hold). I would love to see a candidate GUI that tackles that speed angle.
I don't think we should homage GUI for AI agent workflows. The terminal and Mac GUI are 100% deterministic when you click an icon, but I'm not sure if visualizing agent workflows is the right call.
The problem is that AI workflows are inherently different for each person.
The current approach feels like it's forcing a CLI-based model on users. I also don't think chat is a suitable fundamental unit for task delegation.
I've worked on writing a compiler using both hand-written code and AI. Once the project exceeds a certain size, the chat itself becomes a bottleneck. In those cases, I needed proof and gates like Rocq. So in essence, AI coding requires clear negative gates that should be rejected when approaching the goal.
I think something like Figma's canvas model might become the new interface.
Because no matter how meta you make it, agent workflows are ultimately optimized for the individual. My settings often don't match someone else's.
Computers also have this problem, but in that case, you can enforce defaults. With AI agents, it's different.
That's why I think we need a canvas—something like a workspace next to the user space where agent workflows can be dynamically adjusted.
I focus most of my coding effort on building gates that AI-generated code must pass through when it's being produced. That approach has allowed me to handle much larger volumes of code.
The future agent GUI will be about supervising multiple asynchronous tasks.
In a way, this might end up resembling object-oriented programming—just with task graphs instead of object graphs.
Once AI starts generating code, it flows out like water through a burst dam. It's impossible for any human to fully understand it all. At first, I tried to understand every line, but that actually turned out to be less efficient than just writing the code myself. There's a clear fundamental mismatch between the way AI thinks about code and the way I do.
This mismatch seems like an unsolvable impedance mismatch problem—similar to the one between ORM and SQL. So once you decide to use AI agent code, you have no choice but to shift your focus from reviewing the code itself to trusting the gates you've built around it.
The key question is how tightly you can build those gates. And I don't think chat-based interfaces are capable of providing that level of control.
I only use web chat with copy&paste and file upload option. Anything else I don't trust, especially since the people who create AI harnesses and apps use AI to develop them.
[flagged]
[flagged]
[flagged]
[flagged]
[flagged]
[flagged]
Why are you assuming that the human is in charge? Needing a human to drive the AI is probably a transitional phase.
I'm glad folks are experimenting with different paradigms.
That said, Hacker News might not be the best audience for this informal survey. I'd wager at least 70% of the people reading this are still interacting with their computers using "strange terminal commands", by choice. We aren't representative of what normal people will use or prefer.